Bicycle mass eccentricity is an important performance parameter of unmanned bicycles. In order to reduce the influence of body mass eccentricity on the course trajectory of unmanned bicycles, a probabilistic inference learning optimization (PILO) eccentricity correction method based on the principle of model reinforcement learning was proposed. The method takes the lateral inclination Angle (Angle velocity), handlebar Angle (Angle velocity) and handlebar control moment as inputs, and takes the lateral inclination velocity (Angle acceleration) and handlebar Angle velocity (Angle acceleration) of the vehicle body at the corresponding time as outputs. The probabilistic dynamic model (PDM) of the system is constructed by Gaussian process regression (GPR), which is used for the subsequent state sequence prediction. Taking the mass eccentricity as a parameter of the handlebar PD controller, considering the kinematic constraint between the body heading and the handlebar Angle, the objective function is constructed by the body heading angular velocity, and the mass eccentricity parameter is optimized and corrected. The results show that the absolute error of PILO system is less than 0.005rad, the relative error is less than 10%, and the anti-jamming ability is demonstrated. Compared with the model free cognitive learning eccentric optimization (RLO) calibration system, PILO system has some advantages in parameter setting difficulty, intelligence and fault tolerance.